class DeviceThread;

// Forwards operations to `devices`, maintaining ParallelTensor with components
// placed on each underlying device.
class ParallelDevice {
 public:
  // Eager async execution is only supported when remote eager is not in use
  // (b/157523095).
  explicit ParallelDevice(const std::vector<std::string>& devices,
                          bool is_async = false, int in_flight_nodes_limit = 0);

  ~ParallelDevice();

// A tensorflow.ServerDef specifies remote workers (in addition to the current
// workers name). Operations created in this context can then be executed on
// any of these remote workers by setting an appropriate device.
//
// If the following is set, all servers identified by the
// ServerDef must be up when the context is created.
TF_CAPI_EXPORT extern void TFE_ContextSetServerDef(TFE_Context* ctx,

// gradients available initially).
//
// Then we repeatedly pop an entry from the stack, run its backprop, and update
// the gradients of its inputs. Once we have computed all gradients for a single
// input we can mark this input as done, and this can trigger adding an entry to
// the stack if all outputs of that entry are now done.
//
// When the stack is empty we have gradients for all tensors we're interested
// in.

//
// The API here is subject to changes in the future.
// --------------------------------------------------------------------------

#ifdef __cplusplus
extern "C" {
#endif

// When `enable` is true, set
// tensorflow.ConfigProto.OptimizerOptions.global_jit_level to ON_1, and also
// set XLA flag values to prepare for XLA compilation. Otherwise set
// global_jit_level to OFF.
//

// APIs for generically dealing with op attributes (e.g. when forwarding them
// through custom device implementations).
//
// TODO(allenl): Currently these are black boxes, but we should have some way to
// inspect values. This would let people e.g. copy over most attributes and then
// modify some based on their values.

// A reference to an op's name -> attribute mapping

// nodes. This may cause incorrectness in graph analyses, but is useful for
// simulation or scheduling.
// If aggressive_shape_inference is true, nodes are executed on the host to
// identify output values when possible and does other aggressive strategies.
// This may cause incorrectness in graph analyses, but is useful for simulation
// or scheduling.
// If include_input_tensor_values is true, the values of constant

/// This class should be shared by read-only random access files on a remote
/// filesystem (e.g. GCS).
class RamFileBlockCache {
 public:
  /// The callback executed when a block is not found in the cache, and needs to
  /// be fetched from the backing filesystem. This callback is provided when the
  /// cache is constructed. It returns total bytes read ( -1 in case of errors
  /// ). The `status` should be `TF_OK` as long as the read from the remote

  // TF eager runtime.
  // TODO(b/172877902): These helper functions are currently used to support
  // PyFuncOp on TFRT, and might be useful for ops that directly use low
  // level TF APIs. Remove/replace the following functions when TFRT native
  // ops are implemented.
  //===--------------------------------------------------------------------===//
  // Create an abstract tensor handle from tensorflow::Tensor.